Intelligent calculation task migration-oriented SRv6 multi-path dynamic arrangement method and system

Abstract

The invention provides an intelligent calculation task migration-oriented SRv6 multi-path dynamic arrangement method and system, and the method comprises the steps: carrying out the feature structuring of the obtained intelligent calculation task feature data and network link state data, and generating a multi-dimensional structured data table which comprises the task emergency degree, the network link state data, and the task emergency degree. A target node coordinate and a time delay sensitivity coefficient; performing multi-dimensional path decision on the multi-dimensional structured data table through a multi-objective joint optimization algorithm to generate an optimized path distribution scheme; according to the optimized path distribution scheme, migration channel dynamic control is executed, and a task migration execution log is output; and generating a network configuration optimization instruction based on an index deviation between the task migration execution log and a preset task demand table. By adopting the method, the network path can be hierarchically divided and the proper migration channel can be dynamically selected based on the emergency degree of the task and the position of the target node, so that the migration efficiency and reliability of the intelligent calculation task are improved.

Description

Technical Field

[0001] This invention belongs to the field of network communication technology, and in particular relates to an SRv6 multi-path dynamic orchestration method and system for intelligent computing task migration. Background Technology

[0002] With the development of intelligent computing task migration technology and SRv6 (Segment Routing over IPv6) network technology, SRv6-based multipath transmission technology has emerged. This technology flexibly schedules traffic to multiple paths through source routing mechanisms, aiming to improve network resource utilization and transmission reliability. In traditional implementations, SRv6 multipath transmission mainly relies on static policies or single indicators (such as bandwidth) for path selection, and distributes task traffic through preset path weights or simple load balancing algorithms.

[0003] However, in complex and dynamic network environments, existing methods struggle to simultaneously adapt to real-time changes in multi-dimensional metrics (such as bandwidth, latency, and reliability) and accurately match diverse task requirements. When faced with dynamic factors such as sudden traffic spikes and node heterogeneity, traditional static strategies are prone to causing a disconnect between path selection and task characteristics, leading to risks such as unbalanced resource allocation, decreased transmission efficiency, and even deterioration in service quality. This lack of dynamic coordination capability has become a key bottleneck restricting the migration efficiency of intelligent computing tasks. Summary of the Invention

[0004] Therefore, it is necessary to provide an SRv6 multi-path dynamic orchestration method and system for intelligent computing task migration to address the above-mentioned technical problems. This method can hierarchically divide network paths and dynamically select appropriate migration channels based on the urgency of the task and the location of the target node, thereby improving the efficiency and reliability of intelligent computing task migration.

[0005] Firstly, this application provides an SRv6 multi-path dynamic orchestration method for intelligent computing task migration, including:

[0006] The acquired intelligent computing task feature data and network link status data are structured to generate a multi-dimensional structured data table, which includes task urgency, target node coordinates and latency sensitivity coefficient.

[0007] A multi-objective joint optimization algorithm is used to make multi-dimensional path decisions on multi-dimensional structured data tables and generate an optimized path allocation scheme.

[0008] Based on the optimized path allocation scheme, perform dynamic control of the migration channel and output the task migration execution log;

[0009] Based on the discrepancies between the metrics in the task migration execution log and the preset task requirement table, network configuration optimization instructions are generated.

[0010] In one embodiment, the acquired intelligent computing task feature data and network link status data are structurated to generate a multi-dimensional structured data table, including:

[0011] Semantic parsing of intelligent computing task feature data is performed using a pre-trained domain knowledge graph to extract semantic association rules between task type and transmission requirements, and to generate task requirement mapping vectors.

[0012] Perform spatiotemporal graph convolution operation on network link state data to extract topology change trend features and generate a link state spatiotemporal matrix;

[0013] By integrating the task requirement mapping vector with the link state spatiotemporal matrix, feature dimensionality reduction is performed through dynamic weight perception, and a multi-dimensional structured data table is output.

[0014] In one embodiment, a multi-objective joint optimization algorithm is used to perform multi-dimensional path decision-making on a multi-dimensional structured data table to generate an optimized path allocation scheme, including:

[0015] Based on the latency sensitivity coefficient in the multi-dimensional structured data table, a candidate path set is generated through Pareto front analysis;

[0016] Apply task urgency constraints to the candidate path set to generate a subset of feasible paths;

[0017] Obtain real-time bandwidth availability data for a subset of feasible paths, calculate path hop cost based on target node coordinates, and generate topological distance weights based on path hop cost using the Dijkstra algorithm.

[0018] By integrating real-time bandwidth availability data with topological distance weights, the comprehensive path utility value is calculated using a weighted approximation of the ideal solution ranking method.

[0019] Paths are sorted in descending order of comprehensive utility value, and paths whose utility value reaches a preset threshold are selected to generate an optimized path allocation scheme.

[0020] In one embodiment, based on the deviation between the metrics of the task migration execution log and the preset task requirement table, a network configuration optimization instruction is generated, including:

[0021] Obtain the actual latency data from the task migration execution log;

[0022] Retrieve latency requirement data from the preset task requirement table;

[0023] The actual delay data and the required delay data are processed to calculate the deviation and generate a delay deviation matrix.

[0024] Attribution analysis is performed based on the time delay deviation matrix using the gradient backpropagation algorithm to identify key decision-making defect nodes.

[0025] Adjust the weight distribution parameters of the multi-objective joint optimization algorithm based on the key decision-making defect nodes;

[0026] Generate network configuration optimization instructions that include weight distribution parameters.

[0027] In one embodiment, attribution analysis is performed based on the time delay deviation matrix using a gradient backpropagation algorithm to identify key decision-making defect nodes, including:

[0028] Obtain the decision parameter set of the multi-objective joint optimization algorithm, and construct a forward propagation computation graph containing input layer nodes, hidden layer nodes, and output layer nodes through a directed acyclic graph structure;

[0029] The output layer error value is calculated based on the time delay deviation matrix, and the local gradient of each node in the forward propagation computation graph is calculated layer by layer using the chain rule.

[0030] The local gradient is normalized by absolute value to generate a gradient influence factor vector.

[0031] Nodes whose values ​​in the gradient influence factor vector exceed a dynamic threshold are marked as critical decision-making defect nodes. The dynamic threshold is generated based on historical gradient data using a sliding window statistical method.

[0032] Secondly, this application also provides an SRv6 multi-path dynamic orchestration system for intelligent computing task migration, including:

[0033] The feature structuring module is used to perform feature structuring on the acquired intelligent computing task feature data and network link status data, and generate a multi-dimensional structured data table, which includes task urgency, target node coordinates and latency sensitivity coefficient.

[0034] The path decision module is used to make multi-dimensional path decisions on multi-dimensional structured data tables through a multi-objective joint optimization algorithm, and generate an optimized path allocation scheme.

[0035] The dynamic control module is used to perform dynamic control of the migration channel according to the optimized path allocation scheme and output the task migration execution log.

[0036] The optimization analysis module is used to generate network configuration optimization instructions based on the deviation of indicators between the task migration execution log and the preset task requirement table.

[0037] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned SRv6 multi-path dynamic orchestration method for intelligent computing task migration.

[0038] Fourthly, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-mentioned SRv6 multi-path dynamic orchestration method for intelligent computing task migration.

[0039] The aforementioned SRv6 multi-path dynamic orchestration method and system for intelligent computing task migration utilizes a coherent approach: structuring intelligent computing task feature data and network link status data to generate a multi-dimensional structured data table; employing a multi-objective joint optimization algorithm to perform multi-dimensional path decision-making on the data table to generate an optimized path allocation scheme; executing dynamic control of the migration channel according to this scheme and outputting task migration execution logs; and generating network configuration optimization instructions based on the deviation between the logs and the preset task requirement table. This achieves SRv6 multi-path dynamic orchestration for intelligent computing task migration. This technical solution comprehensively considers multiple factors such as task urgency, target node coordinates, and latency sensitivity, thereby allocating network resources and paths more efficiently, improving the efficiency and accuracy of intelligent computing task migration, ensuring tasks better meet latency requirements, and continuously optimizing network configuration based on actual conditions. This makes the entire intelligent computing task migration process more intelligent, flexible, and highly adaptable, enhancing network performance and reliability in different intelligent computing task migration scenarios. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating an SRv6 multi-path dynamic orchestration method for intelligent computing task migration provided in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the structure of an SRv6 multi-path dynamic orchestration system for intelligent computing task migration provided in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] First, a brief introduction to the terms used in the embodiments of this application will be given.

[0045] Intelligent computing task migration refers to the process of dynamically scheduling and transferring artificial intelligence computing tasks (such as model training, inference services, etc.) and their associated data between heterogeneous computing nodes in a distributed computing environment in order to optimize resource utilization and improve system efficiency. Its core lies in realizing the intelligent redistribution of computing load through network awareness and resource collaboration, thereby ensuring the timeliness of high-priority tasks, improving the overall reliability of the system, and adapting to dynamically changing network conditions and computing needs.

[0046] SRv6 (Segment Routing over IPv6) is a new generation source routing technology based on the IPv6 data plane. It encapsulates instructional routing segments by extending the routing extension header (SRH) in the IPv6 header, enabling network nodes to perform hop-by-hop forwarding according to the path policies preset by the source nodes. This achieves key capabilities such as traffic engineering and fast rerouting while retaining the simplicity of IP technology, providing a programmable, low-latency, and highly reliable transmission service foundation for intelligent computing networks.

[0047] Dijkstra's algorithm is a classic greedy algorithm for finding the single-source shortest path in a weighted directed graph. Its core idea is to gradually determine the minimum cost path from the starting node to all other nodes through iterative traversal. In intelligent computing network path decision-making, this algorithm provides the mathematical basis for optimal path selection for task migration by calculating the quantitative relationship between the topological distance between nodes and the transmission cost, thereby supporting the construction of low-latency and high-reliability network transmission channels.

[0048] Based on the above definitions, the implementation environment of an SRv6 multi-path dynamic orchestration method for intelligent computing task migration provided in this application embodiment will be described. Indicatively, the implementation environment includes: a terminal, sensors, a processor, and a memory. The terminal is connected to the processor, sensors, and memory via a network; sensors include, but are not limited to, network traffic sensors, device status sensors, and environmental perception sensors; the processor can be a central processing unit (CPU), a graphics processing unit (GPU), or an artificial intelligence chip (such as an NPU or TPU); the memory can be a distributed cloud storage system or a local server cluster, without limitation here.

[0049] Based on the above definitions and implementation environment, the application scenarios of the embodiments of this application are described. The SRv6 multi-path dynamic orchestration method for intelligent computing task migration provided in the embodiments of this application can be applied to scenarios including but not limited to the following:

[0050] In cross-regional AI training clusters, model parameter synchronization requires the high-frequency migration of massive gradient data between heterogeneous computing nodes. This technical solution addresses this by real-time sensing of training task stage characteristics (such as the latency-sensitive period of backpropagation), dynamically generating path allocation strategies based on GPU / NPU utilization, and leveraging SRv6's multi-path capabilities to direct key gradient flows to low-latency channels while simultaneously diverting weight update data to high-bandwidth paths. This effectively avoids idle computing nodes and improves distributed training efficiency.

[0051] In the fields of industrial automation and intelligent manufacturing, various intelligent computing tasks in the production process have extremely high requirements for real-time performance and accuracy. This technical solution can be applied to the automated production line network of a factory, dynamically scheduling task migration paths based on the priority of production tasks and equipment status. For example, in an automobile manufacturing workshop, intelligent computing tasks (such as parts inspection and quality control) at different production stages need to be processed promptly. This approach enables efficient scheduling of industrial networks, reduces the risk of production interruptions, improves production efficiency and product quality, and ensures the smooth operation of the intelligent manufacturing process.

[0052] In smart city and IoT applications, a large amount of data generated by devices requires intelligent computing and processing. This technical solution can be applied to scenarios such as urban intelligent transportation systems and intelligent energy management. Taking intelligent transportation as an example, it can dynamically adjust the migration of tasks between edge computing nodes and cloud servers based on traffic flow conditions and the urgency of the task. This helps to achieve functions such as real-time optimization of traffic signals and road condition prediction, improving urban traffic operation efficiency and energy utilization efficiency, and providing strong support for smart city construction.

[0053] As an illustration, the SRv6 multi-path dynamic orchestration method for intelligent computing task migration provided in this application embodiment can also be applied to other application scenarios. This is only an example and does not limit the specific application scenarios.

[0054] In one exemplary embodiment, such as Figure 1 As shown, an SRv6 multi-path dynamic orchestration method for intelligent computing task migration is provided. This embodiment illustrates the application of this method to a terminal in the aforementioned implementation environment. It is understood that this method can also be applied to a server, and also to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 104:

[0055] Step 101: The acquired intelligent computing task feature data and network link status data are structured to generate a multi-dimensional structured data table, which includes the task urgency, target node coordinates, and latency sensitivity coefficient.

[0056] For example, the collection of intelligent computing task characteristic data includes information such as task urgency, target node coordinates, and latency sensitivity coefficient. This data is provided by user input from terminal devices, historical task records, and the intelligent computing task scheduling system. Network link status data is collected through network link monitoring tools, router and switch management interfaces, and network performance analysis software, primarily collecting key indicators such as network bandwidth, latency, and packet loss rate. To ensure data accuracy and completeness, the system preprocesses the collected raw data, including data cleaning to remove errors and duplicate information, data transformation to unify different data formats, and data normalization to ensure data comparison under the same dimensions. The preprocessed data is organized into a multi-dimensional structured data table according to a predetermined data structure, where each dimension corresponds to a specific feature or attribute, such as the task urgency dimension, target node coordinate dimension, and latency sensitivity coefficient dimension.

[0057] Step 102: Perform multi-dimensional path decision-making on the multi-dimensional structured data table using a multi-objective joint optimization algorithm to generate an optimized path allocation scheme.

[0058] For example, a multi-objective joint optimization algorithm comprehensively considers factors such as task urgency, target node location, latency requirements, and network link bandwidth, latency, and stability. Through complex mathematical modeling and algorithmic calculations, it generates an optimized path allocation scheme. For instance, the algorithm may employ heuristic search methods such as genetic algorithms, simulated annealing algorithms, or ant colony algorithms, which can find near-optimal solutions in complex search spaces. For example, genetic algorithms iterate the population to optimize path allocation schemes by simulating natural selection and genetic mechanisms; simulated annealing algorithms gradually converge to the global optimum by simulating the physical annealing process, avoiding the trap of local optima; and ant colony algorithms simulate ant foraging behavior and use pheromone update mechanisms to find the optimal path. Furthermore, when making path decisions, the algorithm assigns a comprehensive weight to each intelligent computing task. This weight is dynamically calculated based on factors such as task urgency and latency sensitivity, ensuring that high-priority tasks obtain better migration paths. In this way, the multi-objective joint optimization algorithm can maximize the utilization of network resources and improve the overall network performance and efficiency while meeting diverse task requirements.

[0059] Step 103: Based on the optimized path allocation scheme, perform dynamic control of the migration channel and output the task migration execution log.

[0060] Specifically, based on the generated optimized path allocation scheme, the migration channel is adjusted and optimized in real time through network control protocols and commands, such as the Software-Defined Networking (SDN) controller and SRv6-related routing protocol extensions. For example, SRv6 tunnels or paths can be dynamically established and maintained according to the optimized path scheme to ensure that intelligent computing tasks migrate along the optimal path. During the migration process, the task migration status and network performance indicators are monitored in real time, and key information during the migration process, such as migration start time, end time, migration path, task data volume, and abnormal situations that occur during the migration, is recorded to form a detailed task migration execution log. Furthermore, based on real-time monitored changes in network status, the parameters of the migration channel, such as bandwidth allocation and path switching, can be dynamically adjusted to cope with sudden network situations and ensure the stability and reliability of task migration. In this way, not only can efficient task migration be achieved, but problems during the migration process can also be detected and handled in a timely manner, providing strong data support for subsequent network optimization.

[0061] Step 104: Generate network configuration optimization instructions based on the deviation between the indicators in the task migration execution log and the preset task requirement table.

[0062] Specifically, the actual migration metrics (such as migration time, latency, and data integrity) in the execution log are compared and analyzed with the expected metrics in the preset task requirement table to calculate the metric deviation. For example, if the preset task requirement table specifies that the migration latency of a certain task should be less than 100 milliseconds, while the actual execution log shows 120 milliseconds, then the latency deviation is 20 milliseconds. Exemplarily, based on these deviations, deviation analysis algorithms and network configuration optimization models, such as machine learning-based prediction models or rule-based inference models, are used to generate corresponding network configuration optimization instructions. Furthermore, these instructions may involve adjusting network device parameters, optimizing routing strategies, increasing bandwidth resources, or improving network topology to reduce deviations and improve the performance and quality of subsequent task migrations. For example, based on latency deviations, instructions to increase the bandwidth of specific links or adjust routes can be generated to reduce latency; or based on data integrity deviations, data transmission protocols can be optimized or redundancy mechanisms can be added to improve data transmission reliability. In this way, this technical method can achieve continuous optimization of network configuration, continuously improving network performance and adaptability to better meet the diverse needs of intelligent computing tasks.

[0063] The aforementioned SRv6 multi-path dynamic orchestration method for intelligent computing task migration utilizes a coherent approach: structuring intelligent computing task feature data and network link status data to generate a multi-dimensional structured data table; employing a multi-objective joint optimization algorithm to perform multi-dimensional path decision-making on the data table to generate an optimized path allocation scheme; executing dynamic control of the migration channel according to this scheme and outputting task migration execution logs; and generating network configuration optimization instructions based on the deviation between the logs and the preset task requirement table. This method achieves SRv6 multi-path dynamic orchestration for intelligent computing task migration. This technical solution comprehensively considers multiple factors such as task urgency, target node coordinates, and latency sensitivity, thereby allocating network resources and paths more efficiently, improving the efficiency and accuracy of intelligent computing task migration, ensuring tasks better meet latency requirements, and continuously optimizing network configuration based on actual conditions. This makes the entire intelligent computing task migration process more intelligent, flexible, and highly adaptable, enhancing network performance and reliability in different intelligent computing task migration scenarios.

[0064] In one embodiment, the acquired intelligent computing task feature data and network link status data are structurated to generate a multi-dimensional structured data table, including:

[0065] Semantic parsing of intelligent computing task feature data is performed using a pre-trained domain knowledge graph to extract semantic association rules between task type and transmission requirements, and to generate task requirement mapping vectors.

[0066] Specifically, semantic parsing of intelligent computing task feature data is performed using a pre-trained domain knowledge graph. Intelligent computing task feature data typically originates from user requests, application systems, or historical task records, and includes basic task characteristics such as task type, expected completion time, and data volume. The pre-trained domain knowledge graph is trained on a large amount of text data from a specific domain and is capable of understanding the semantic information in the task feature data. For example, in the cloud computing domain, the knowledge graph may contain entities such as various cloud service types, resource requirements, and performance metrics, as well as their interrelationships. By inputting the task feature data into the domain knowledge graph, this technical solution can extract semantic association rules between task type and transmission requirements, thereby generating a task requirement mapping vector. This process can be implemented using a graph neural network (GNN) algorithm, where GNNs can capture complex relationships in graph structures and transform task features into low-dimensional mapping vector representations.

[0067] Perform spatiotemporal graph convolution operation on network link state data to extract topology change trend features and generate a link state spatiotemporal matrix.

[0068] Specifically, network link-state data is typically provided by network infrastructure such as network monitoring devices, routers, and switches, and includes metrics such as bandwidth utilization, latency, and packet loss rate. Spatiotemporal graph convolution operations combine Graph Convolutional Networks (GCNs) and time series analysis methods, simultaneously considering the spatial structure and temporal variation characteristics of the network topology. For example, an adjacency matrix is ​​constructed to represent the connections between network nodes, and the link-state data of each node is used as a node feature vector. Graph convolution operations aggregate information from neighboring nodes in the spatial dimension, and time series models (such as LSTM or GRU) are combined to capture the dynamic trends of link-state changes in the temporal dimension. After the spatiotemporal graph convolution operation, a link-state spatiotemporal matrix is ​​generated. This matrix stores the spatiotemporal characteristics of the network link-state in matrix form, providing rich topological and temporal information for subsequent path decisions.

[0069] By integrating the task requirement mapping vector with the link state spatiotemporal matrix, feature dimensionality reduction is performed through dynamic weight perception, and a multi-dimensional structured data table is output.

[0070] Specifically, the task requirement mapping vector and the link state spatiotemporal matrix can be fused through concatenation, element-wise multiplication, or other feature fusion methods. For example, the task requirement mapping vector can be used as an enhancement of node features and concatenated with node features in the link state spatiotemporal matrix to form a fused feature matrix. A dynamic weight-aware mechanism is then used to reduce the dimensionality of the fused features. This mechanism dynamically adjusts the weights of each feature based on the current network state and task requirements, thus preserving the most valuable information for path decision-making. An attention mechanism can be employed, where attention weights are adjusted dynamically based on changes in the input features. The multi-dimensional structured data table output by this approach not only contains key task requirement information but also incorporates real-time network link state features, providing comprehensive, accurate, and efficient data support for subsequent multi-objective joint optimization algorithms. This embodiment combines semantic parsing of domain knowledge graphs with spatiotemporal graph convolution operations to deeply mine potential information in task feature data and network link state data, improving the accuracy and richness of feature representation. At the same time, the dynamic weight-aware feature dimensionality reduction method can reduce data dimensionality while retaining key information, thereby improving data processing efficiency and the performance of subsequent path decision-making.

[0071] In one embodiment, a multi-objective joint optimization algorithm is used to perform multi-dimensional path decision-making on a multi-dimensional structured data table to generate an optimized path allocation scheme, including:

[0072] Based on the latency sensitivity coefficient in the multi-dimensional structured data table, a candidate path set is generated through Pareto front analysis;

[0073] Apply task urgency constraints to the candidate path set to generate a subset of feasible paths;

[0074] Obtain real-time bandwidth availability data for a subset of feasible paths, calculate path hop cost based on target node coordinates, and generate topological distance weights based on path hop cost using the Dijkstra algorithm.

[0075] By integrating real-time bandwidth availability data with topological distance weights, the comprehensive path utility value is calculated using a weighted approximation of the ideal solution ranking method.

[0076] Paths are sorted in descending order of comprehensive utility value, and paths whose utility value reaches a preset threshold are selected to generate an optimized path allocation scheme.

[0077] For example, Pareto front analysis is used to generate a candidate path set based on latency sensitivity coefficients in a multi-dimensional structured data table. Pareto front analysis is a method for multi-objective optimization decision-making that identifies optimal path combinations by finding solutions that cannot be further improved among multiple objectives. For example, for tasks with high latency sensitivity, Pareto front analysis prioritizes paths that perform best in key metrics such as latency and bandwidth, thus forming a candidate path set containing multiple potentially optimal paths. Task urgency constraints are imposed on the generated candidate path set to generate a subset of feasible paths. Task urgency constraints come from a task scheduling system or user-defined priority rules. For example, for tasks with high urgency, this method filters out paths that meet low latency and high bandwidth requirements, thereby narrowing down the range of feasible paths. Specifically, this can be achieved by setting an urgency threshold; only paths that meet or exceed this threshold are retained in the subset of feasible paths. Further, real-time bandwidth availability data for the subset of feasible paths is obtained, and the path hop cost is calculated based on the target node coordinates. Real-time bandwidth availability data is typically provided by network monitoring systems, reflecting the current available bandwidth of network links. Path hop cost is calculated based on the number of hops between nodes in the path; fewer hops generally mean a shorter path and higher transmission efficiency. For example, the coordinates of the target node can be obtained from a network topology database, and the hop count of each path can be calculated using graph theory algorithms. Based on this, the topological distance weights are calculated using the Dijkstra algorithm. The Dijkstra algorithm is a classic shortest path algorithm used to find the shortest path between two points in a graph. By combining real-time bandwidth availability data and topological distance weights, the overall path utility value is calculated using the Weighted Approximation Ideal Solution Ranking Method (TOPSIS). TOPSIS is a commonly used multi-attribute decision analysis method that can comprehensively consider multiple indicators to rank alternative solutions. For example, real-time bandwidth availability data and topological distance weights are used as two key indicators, each assigned an appropriate weight, and the overall utility value of each path is calculated. The paths are then sorted in descending order based on the calculated overall utility values, and paths with utility values ​​reaching a preset threshold are selected to generate an optimized path allocation scheme.

[0078] This embodiment, by combining Pareto front analysis, urgency constraints, real-time bandwidth monitoring, path hop count calculation, and weighted ranking methods, comprehensively considers the latency sensitivity, urgency, and real-time status of network resources of the task, thereby generating an efficient and adaptable path allocation scheme. This method not only improves the accuracy and reliability of path decisions but also enhances the utilization rate of network resources and the success rate of task migration, providing strong support for the migration of intelligent computing tasks in complex network environments.

[0079] In one embodiment, based on the deviation between the metrics of the task migration execution log and the preset task requirement table, a network configuration optimization instruction is generated, including:

[0080] Obtain the actual latency data from the task migration execution log;

[0081] Retrieve latency requirement data from the preset task requirement table;

[0082] The actual delay data and the required delay data are processed to calculate the deviation and generate a delay deviation matrix.

[0083] Attribution analysis is performed based on the time delay deviation matrix using the gradient backpropagation algorithm to identify key decision-making defect nodes.

[0084] Adjust the weight distribution parameters of the multi-objective joint optimization algorithm based on the key decision-making defect nodes;

[0085] Generate network configuration optimization instructions that include weight distribution parameters.

[0086] Specifically, the actual latency data reflects the actual time consumed during the task migration process. The preset task requirement table is predefined based on the specific needs of the intelligent computing task, including requirements for latency, bandwidth, data integrity, and other aspects. Based on the latency deviation matrix, this technical solution uses the gradient backpropagation algorithm for attribution analysis to identify key decision-making defect nodes. The gradient backpropagation algorithm is an optimization algorithm originating from the field of machine learning, which can effectively track the source of error. This method can accurately locate key decision-making defect nodes, providing a basis for subsequent algorithm parameter adjustments. Based on the results of the attribution analysis, the weight distribution in the multi-objective joint optimization algorithm is redistributed. For example, if the processing latency of a specific node is found to be the main cause of latency deviation, the weight of that node in the network path decision can be appropriately reduced, while the weights of other nodes or links can be increased to optimize path selection. This technical solution generates network configuration optimization instructions containing weight distribution parameters to directly guide network devices to adjust their configurations, optimize network resource allocation, and improve task migration performance. This embodiment achieves intelligent optimization of network configuration by combining deviation analysis of task migration execution logs and a preset task requirement table, utilizing the gradient backpropagation algorithm for accurate attribution, and dynamically adjusting the weight distribution of a multi-objective joint optimization algorithm. This method not only effectively reduces latency deviations during task migration but also improves network resource utilization efficiency, ensuring efficient migration and execution of intelligent computing tasks.

[0087] In one embodiment, attribution analysis is performed based on the time delay deviation matrix using a gradient backpropagation algorithm to identify key decision-making defect nodes, including:

[0088] Obtain the decision parameter set of the multi-objective joint optimization algorithm, and construct a forward propagation computation graph containing input layer nodes, hidden layer nodes, and output layer nodes through a directed acyclic graph structure;

[0089] The output layer error value is calculated based on the time delay deviation matrix, and the local gradient of each node in the forward propagation computation graph is calculated layer by layer using the chain rule.

[0090] The local gradient is normalized by absolute value to generate a gradient influence factor vector.

[0091] Nodes whose values ​​in the gradient influence factor vector exceed a dynamic threshold are marked as critical decision-making defect nodes. The dynamic threshold is generated based on historical gradient data using a sliding window statistical method.

[0092] For example, a set of decision parameters for a multi-objective joint optimization algorithm is obtained, including parameters such as weights and biases involved in the algorithm. These parameters reflect the degree of importance the algorithm places on different features during path decision-making. A forward propagation computation graph containing input layer nodes, hidden layer nodes, and output layer nodes is constructed using a directed acyclic graph (DAG) structure. DAG is a common deep learning model representation method that clearly shows the flow direction and processing of data between different layers. For example, input layer nodes may receive feature data such as task urgency and latency sensitivity coefficients from a multi-dimensional structured data table; hidden layer nodes perform nonlinear combinations and transformations of features; and output layer nodes generate the final path decision result. The latency bias matrix stores the deviation information between the actual and required latency of task migration, and this information is input back into the computation graph as an error signal. The chain rule is a mathematical tool used to calculate the derivative of a composite function; it can calculate the contribution of each node to the final error layer by layer, i.e., the local gradient. For example, for output layer nodes, their local gradients can be obtained by calculating the derivative of the time delay bias matrix with the output result. Then, the local gradients of hidden layer and input layer nodes are calculated sequentially until the gradients of all nodes are calculated. Normalization aims to eliminate the influence of dimensional differences between different features, making gradient values ​​comparable. Absolute value normalization can employ methods such as max-min normalization and Z-score normalization to transform the absolute value of the local gradient to a fixed range (e.g., [0,1]), generating a gradient influence factor vector. Each element in this vector represents the degree of influence of the corresponding node in path decision-making. The sliding window statistical method is a data processing technique that selects historical gradient data from a recent period as samples, calculates its statistical characteristics (e.g., mean, standard deviation), and dynamically determines a threshold based on these characteristics. For example, a dynamic threshold can be set to twice the standard deviation of the historical gradient data mean. Nodes with gradient influence factors exceeding this threshold are considered critical decision-making defect nodes, which may be the main cause of time delay bias. By identifying these key nodes, this technical solution can provide precise location and guidance for subsequent algorithm optimization and network configuration adjustments, thereby effectively improving the latency performance of task migration.

[0093] This embodiment, by combining the gradient backpropagation algorithm and the dynamic threshold labeling method, can accurately identify key decision-making defects in multi-objective joint optimization algorithms, providing data-driven decision-making basis for algorithm parameter optimization and network configuration adjustment. The above method not only improves the accuracy and efficiency of latency deviation attribution analysis but also enhances the adaptability and optimization effect of network configuration, contributing to improving the overall performance and reliability of intelligent computing task migration.

[0094] In summary, the SRv6 multi-path dynamic orchestration method for intelligent computing task migration provided in this application generates a multi-dimensional data table containing task urgency, target node coordinates, and latency sensitivity coefficients by jointly structuring intelligent computing task feature data and network link status data, thus establishing a quantitative correlation between tasks and the network. Based on the latency sensitivity coefficients, a candidate path set is generated through Pareto front analysis. A subset of feasible paths is selected by combining task urgency constraints. The topological distance weights calculated from the target node coordinates and real-time bandwidth availability data are integrated, and a weighted approximation of the ideal solution ranking method is used to generate a comprehensive path utility value. Based on this, the optimal path is dynamically selected to form an allocation scheme. SRv6 multi-path dynamic control is executed, and traffic forwarding weights are adjusted in real-time based on reinforcement learning, and migration logs are output. The latency deviation between the logs and the demand table is analyzed using the gradient backpropagation algorithm to locate decision-making defective nodes and optimize weight parameters in a closed loop. This technical solution builds a quantitative decision-making foundation by deeply integrating task characteristics and network status modeling, and adopts a multi-level joint optimization architecture to achieve dynamic hierarchical filtering of paths, thereby improving the efficiency and reliability of intelligent computing task migration and ensuring that intelligent computing tasks can migrate efficiently and stably in complex and ever-changing network environments.

[0095] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0096] Based on the same inventive concept, this application also provides an SRv6 multi-path dynamic orchestration system 10 for implementing the aforementioned SRv6 multi-path dynamic orchestration method for intelligent computing task migration. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the SRv6 multi-path dynamic orchestration system 10 for intelligent computing task migration provided below can be found in the limitations of the SRv6 multi-path dynamic orchestration method for intelligent computing task migration described above, and will not be repeated here.

[0097] In one exemplary embodiment, such as Figure 2 As shown, an SRv6 multi-path dynamic orchestration system 10 for intelligent computing task migration is provided, including:

[0098] The feature structuring module 11 is used to perform feature structuring on the acquired intelligent computing task feature data and network link status data to generate a multi-dimensional structured data table, wherein the multi-dimensional structured data table includes task urgency, target node coordinates and latency sensitivity coefficient.

[0099] Path decision module 12 is used to make multi-dimensional path decisions on multi-dimensional structured data tables through a multi-objective joint optimization algorithm and generate an optimized path allocation scheme.

[0100] Dynamic control module 13 is used to perform dynamic control of the migration channel according to the optimized path allocation scheme and output the task migration execution log;

[0101] The optimization analysis module 14 is used to generate network configuration optimization instructions based on the deviation of indicators between the task migration execution log and the preset task requirement table.

[0102] In one embodiment, the feature structuring module 11 includes:

[0103] The semantic parsing unit is used to perform semantic parsing on the feature data of intelligent computing tasks through a pre-trained domain knowledge graph, extract semantic association rules between task types and transmission requirements, and generate task requirement mapping vectors.

[0104] The graph convolution unit is used to perform spatiotemporal graph convolution operations on network link state data, extract topology change trend features, and generate a link state spatiotemporal matrix.

[0105] The feature fusion unit is used to fuse the task requirement mapping vector and the link state spatiotemporal matrix, perform feature dimensionality reduction through dynamic weight perception, and output a multi-dimensional structured data table.

[0106] In one embodiment, the path decision module 12 includes:

[0107] The Pareto analysis unit is used to generate a set of candidate paths through Pareto front analysis based on the time-sensitivity coefficients in a multi-dimensional structured data table.

[0108] The constraint filtering unit is used to impose task urgency constraints on the candidate path set and generate a subset of feasible paths.

[0109] The weight calculation unit is used to obtain real-time bandwidth availability data of feasible path subsets, calculate path hop cost based on target node coordinates, and generate topological distance weights based on path hop cost using Dijkstra's algorithm.

[0110] The utility decision unit is used to integrate real-time bandwidth availability data with topological distance weights and calculate the comprehensive path utility value through a weighted approximation of the ideal solution ranking method.

[0111] The path filtering unit is used to sort paths in descending order based on their comprehensive utility value, select paths whose utility value reaches a preset threshold, and generate an optimized path allocation scheme.

[0112] In one embodiment, the optimization analysis module 14 includes:

[0113] The actual latency acquisition unit is used to acquire the actual latency data in the task migration execution log;

[0114] The latency acquisition unit is used to acquire latency requirement data from the preset task requirement table.

[0115] The deviation calculation unit is used to perform deviation calculation processing on the actual delay data and the delay requirement data to generate a delay deviation matrix;

[0116] The attribution analysis unit is used to perform attribution analysis based on the time delay deviation matrix using the gradient backpropagation algorithm to identify key decision-making defect nodes.

[0117] The weight adjustment unit is used to adjust the weight distribution parameters of the multi-objective joint optimization algorithm according to the key decision-making defect nodes.

[0118] The instruction generation unit is used to generate network configuration optimization instructions that include weight distribution parameters.

[0119] In one embodiment, the attribution analysis unit includes:

[0120] The computation graph construction subunit is used to obtain the decision parameter set of the multi-objective joint optimization algorithm. It constructs a forward propagation computation graph containing input layer nodes, hidden layer nodes, and output layer nodes through a directed acyclic graph structure.

[0121] The gradient calculation subunit is used to calculate the output layer error value based on the time delay deviation matrix, and calculates the local gradient of each node in the forward propagation computation graph layer by layer using the chain rule.

[0122] The normalization sub-unit is used to perform absolute value normalization on the local gradient and generate a gradient influence factor vector.

[0123] The dynamic threshold generation subunit is used to mark nodes in the gradient influence factor vector whose values ​​exceed the dynamic threshold as key decision-making defect nodes. The dynamic threshold is generated based on historical gradient data using the sliding window statistical method.

[0124] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the SRv6 multipath dynamic orchestration method for intelligent computing task migration as described above.

[0125] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0126] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0127] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An SRv6 multi-path dynamic orchestration method for intelligent computing task migration, characterized in that, The method includes: The acquired intelligent computing task feature data and network link status data are structured to generate a multi-dimensional structured data table, wherein the multi-dimensional structured data table includes task urgency, target node coordinates and latency sensitivity coefficient. A multi-objective joint optimization algorithm is used to make multi-dimensional path decisions on the multi-dimensional structured data table to generate an optimized path allocation scheme. Based on the optimized path allocation scheme, perform dynamic control of the migration channel and output the task migration execution log; Based on the deviation between the indicators in the task migration execution log and the preset task requirement table, a network configuration optimization instruction is generated.

2. The method according to claim 1, characterized in that, The process of structuring the acquired intelligent computing task feature data and network link status data to generate a multi-dimensional structured data table includes: Semantic parsing of the intelligent computing task feature data is performed using a pre-trained domain knowledge graph to extract semantic association rules between task type and transmission requirements, and to generate a task requirement mapping vector. Perform a spatiotemporal graph convolution operation on the network link state data to extract topology change trend features and generate a link state spatiotemporal matrix; By integrating the task requirement mapping vector with the link state spatiotemporal matrix, feature dimensionality reduction is performed through dynamic weight perception, and the multi-dimensional structured data table is output.

3. The method according to claim 1, characterized in that, The step of performing multi-dimensional path decision-making on the multi-dimensional structured data table using a multi-objective joint optimization algorithm to generate an optimized path allocation scheme includes: Based on the time-delay sensitivity coefficient in the multi-dimensional structured data table, a candidate path set is generated through Pareto front analysis; Apply task urgency constraints to the candidate path set to generate a subset of feasible paths; Obtain real-time bandwidth availability data for the feasible path subset, calculate path hop cost based on the target node coordinates, and generate topological distance weights based on the path hop cost using the Dijkstra algorithm. By integrating the real-time bandwidth availability data with the topological distance weights, the comprehensive path utility value is calculated using a weighted approximation ideal solution sorting method. The paths are sorted in descending order based on the comprehensive utility value, and the optimized path allocation scheme is generated by selecting the paths whose utility values ​​reach a preset threshold.

4. The method according to claim 1, characterized in that, The step of generating network configuration optimization instructions based on the deviation between the indicators in the task migration execution log and the preset task requirement table includes: Obtain the actual latency data from the task migration execution log; Obtain the latency requirement data from the preset task requirement table; The actual delay data and the delay requirement data are processed to calculate the deviation and generate a delay deviation matrix; Based on the aforementioned time delay deviation matrix, attribution analysis is performed using the gradient backpropagation algorithm to identify key decision-making defect nodes. Adjust the weight distribution parameters of the multi-objective joint optimization algorithm according to the key decision defect nodes; Generate the network configuration optimization instructions that include the weight distribution parameters.

5. The method according to any one of claims 1 to 4, characterized in that, The attribution analysis based on the time delay deviation matrix using the gradient backpropagation algorithm to identify key decision-making defect nodes includes: Obtain the decision parameter set of the multi-objective joint optimization algorithm, and construct a forward propagation computation graph containing input layer nodes, hidden layer nodes, and output layer nodes through a directed acyclic graph structure; The output layer error value is calculated based on the time delay deviation matrix, and the local gradient of each node in the forward propagation computation graph is calculated layer by layer using the chain rule. The local gradient is normalized by absolute value to generate a gradient influence factor vector. Nodes in the gradient influence factor vector whose values ​​exceed a dynamic threshold are marked as critical decision-making defect nodes, wherein the dynamic threshold is generated based on historical gradient data using a sliding window statistical method.

6. An SRv6 multi-path dynamic orchestration system for intelligent computing task migration, characterized in that, The system includes: The feature structuring module is used to perform feature structuring on the acquired intelligent computing task feature data and network link status data to generate a multi-dimensional structured data table, wherein the multi-dimensional structured data table includes task urgency, target node coordinates and latency sensitivity coefficient. The path decision module is used to perform multi-dimensional path decision-making on the multi-dimensional structured data table through a multi-objective joint optimization algorithm, and generate an optimized path allocation scheme. The dynamic control module is used to perform dynamic control of the migration channel according to the optimized path allocation scheme and output the task migration execution log. The optimization analysis module is used to generate network configuration optimization instructions based on the deviation of indicators between the task migration execution log and the preset task requirement table.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images